基于残差网络与注意力机制的脉象信号分析识别

Wrist pulse analysis and recognition based on residual network and attention mechanism

传统机器学习方法在中医脉象信号分类研究中需要人工提取脉象的特征,且分类准确率不高,因此较难应用于实际环境中。为对脉象信号进行特征的自动提取和精确分类,文中采用改进的一维残差网络对脉象信号进行特征提取并结合改进的注意力机制提高网络的分类性能。所构建的基于一维残差网络与注意力机制模型在单个GTX 1080 GPU上对脉象数据集进行实验,得到的平均准确率、平均召回率、平均精确率、平均F1分数均高于98.84%,在临床病例数据集中,该模型对冠心病分类的平均准确率为100%、平均召回率为99.91%、平均精确率为99.54%、平均F1分数为99.72%。与已有的脉象信号分类模型相比,该模型在脉象数据集上和临床病例数据集上的准确率分别提高1.32%~9.23%和0.21%~3.03%。存储该模型时所需容量不超过10 MB,有利于将其部署在实际脉诊设备上并应用于现实的脉象检测。

Traditional machine learning methods need to extract pulse characteristics manually in pulse classification research,and their classification accuracy is not high,so it is difficult to apply in the actual environment.In order to automatically extract and classify the pulse signal features,a one⁃dimensional residual network is used to extract the features of the pulse signal and the attention mechanism is combined to improve the classification performance of the network.The constructed model based on improved one⁃dimensional residual network and improved attention mechanism was used to test the pulse dataset on a single GTX 1080 GPU,and the average accuracy,average recall rate,average precision rate and average F1 score were all higher than 98.84%.The clinical dataset indicates that the average accuracy of the model for coronary heart disease classification is 100%,its average recall rate is 99.91%,its average precision rate is 99.54%,and its average F1 score is 99.72%.In comparison with the existing pulse classification models,the accuracy of this model on the pulse dataset and the clinical dataset is 1.32%~9.23%and 0.21%~3.03%higher,respectively.When storing the model,the required capacity does not exceed 10 MB,which is beneficial to deploy it on the actual pulse diagnosis equipment to realize the pulse detection.